You are on page 1of 29

ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4

SVANTE JANSON
arXiv:1009.2373v1 [math.HO] 13 Sep 2010

Abstract. We present the solutions of equations of degrees 3 and 4


using Galois theory and some simple Fourier analysis for finite groups.

1. Introduction
The purpose of this note is to present the solutions of equations of degrees
3 and 4 (a.k.a. cubic and quartic equations) in a way connected to Galois
theory. This is, of course, not the historical path; the solutions were found
by del Ferro, Tartaglia, Cardano (Cardan) and Ferrari in the 16th century,
see Appendix B, about 300 years before Galois theory was created. But in
retrospect, Galois theory yields the solutions rather easily. More precisely,
we will see below that Galois theory, together with some simple Fourier
analysis for (small) finite Abelian groups, suggests the crucial constructions
in the solutions; however, all calculations are easily verified directly, and
we do not really need any results from Galois theory (or Fourier analysis)
for the solution. Nevertheless, we find it instructive to use Galois theory
as much as possible in order to motivate the constructions. (See also [21,
Section 8.8] for a similar, but not identical, treatment.) The Galois theory
used here can be found in e.g. [2], [4], [8], [9] or [21].
The appendices contain comments on the history of the problem, other
solutions, and (Appendix A) the complications that may arise when we work
with real numbers instead of complex.
Remark 1.1. In contrast, full use of Galois theory is needed to show the
impossibility of similar formulas for solutions of equations of degree 5 or
more. This will not be discussed here; see instead e.g. [2], [4], [8], [9] or [21].
We let throughout K be a field with characteristic 0. (Actually, every-
thing in this note is valid also for a field K of positive characteristic p 6= 2, 3.
However, the cases when the characteristic is 2 or 3 are different since we
divide by 2 and 3 in the formulas below; there are also problems with sepa-
rability in these cases.)
The roots of a polynomial in K[x] are, in general, not elements of K, so we
will work in some unspecified extension of K. This extension could be the
algebraic closure K of K or some other algebraically closed field containing
K; in particular, if K = Q or another subfield of C (as the reader may
assume for simplicity), we can work in C.
Date: 1 December 2009; revised 17 August 2010.
1
2 SVANTE JANSON

For simplicity we consider monic polynomials only. (It is trivial to reduce


to this case by dividing by the leading coefficient.)

2. Polynomials of degree 3
Let f (x) = x3 + bx2 + cx + d, with b, c, d ∈ K, be a polynomial of degree
3, and let α1 , α2 , α3 be its roots in some extension of K. Thus
f (x) = x3 + bx2 + cx + d = (x − α1 )(x − α2 )(x − α3 ). (2.1)
It is convenient to make the translation x = y − b/3, converting f (x) into
g(y) := f (y − b/3) = y 3 + py + q (2.2)
(without second degree term) for some p, q ∈ K. (Such polynomials, without
the second highest degree term, are called reduced or sometimes depressed.)
Thus g has the roots β1 , β2 , β3 with βi = αi + b/3, so αi = βi − b/3, i =
1, . . . , 3. Hence,
g(y) = y 3 + py + q = (y − β1 )(y − β2 )(y − β3 ). (2.3)
Consequently, identifying coefficients,
β1 + β2 + β3 = 0, (2.4)
β1 β2 + β1 β3 + β2 β3 = p, (2.5)
β1 β2 β3 = −q, (2.6)
Remark 2.1. Explicitly,
p = c − 31 b2 , (2.7)
1 2 3
q =d− 3 bc + 27 b . (2.8)
The polynomials f and g have the same discriminant
Y Y
∆ := Dis(f ) = Dis(g) := (αi − αj )2 = (βi − βj )2 . (2.9)
1≤i<j≤3 1≤i<j≤3

Since ∆ is a symmetric polynomial in α1 , α2 , α3 (or β1 , β2 , β3 ), it can be


written as a polynomial in the coefficients of f or g. A well-known calculation
yields, see e.g. [14],
∆ = b2 c2 − 4c3 − 4b3 d + 18bcd − 27d2 = −4p3 − 27q 2 . (2.10)
We also define the square root of ∆:
Y Y √
δ := (αi − αj ) = (βi − βj ) = ∆. (2.11)
1≤i<j≤3 1≤i<j≤3

Note that while ∆ is independent of the ordering of the roots, the sign of δ
may change if we permute α1 , α2 , α3 . More precisely, the sign is preserved
by an even permutation but is changed by an odd permutation.
Let E = K(α1 , α2 , α3 ) = K(β1 , β2 , β3 ) be the splitting field of f , or g, over
K, and let G := Gal(E : K) be the Galois group of the extension E ⊇ K.
The elements of the Galois group G permute the roots αi (or βi ), and G
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 3

may be regarded as a subgroup of the symmetric group S3 . As said above,


σ(δ) = δ if σ ∈ G is an even permutation, while σ(δ) = −δ if σ is odd.
Since K ⊆ K(δ) ⊆ E, E is a Galois extension of K(δ), and the Galois group
Gal(E : K(δ)) is the subgroup of G fixing δ:
Gal(E : K(δ)) = {σ ∈ G : σ(δ) = δ} = {σ ∈ G : σ is even} = G∩A3 (2.12)
(where A3 is the alternating group consisting of all even permutations in S3 ),
at least if δ 6= 0, or equivalently ∆ 6= 0, which is equivalent to f separable
(i.e., f has no multiple roots in K). In particular, if f is irreducible, in which
case G is transitive and thus G = S3 or A3 , Gal(E : K(δ)) = A3 , which is the
cyclic group C3 . Moreover, in this case, Gal(E : K(δ)) = A3 ∼ = C3 acts on
the vectors (α1 , α2 , α3 ) and (β1 , β2 , β3 ) by cyclic permutations; equivalently,
if we regard these vectors as functions on Z3 ∼ = C3 , Gal(E : K(δ)) ∼
= C3 acts
by translations in C3 . This suggests using Fourier analysis, or equivalently
group representation theory, for A3 ∼ = C3 . (For Fourier analysis on finite
Abelian groups, see e.g. [7]; in this case, the Fourier transform is often
called the discrete Fourier transform. The more complicated theory of group
representations for general finite groups is treated by [20].)
Remark 2.2. The method below was given by Lagrange [17] in 1770–1771,
as part of his study of equations of higher degree, see [4, Sections 8.3, 12.1
and p. 14]. The method is thus some decades older than both Galois theory
and Fourier analysis. In this context, the Fourier transforms u and v in
(2.17)–(2.18) below (or rather 3u and 3v) are known as the Lagrange resol-
vents for the equation. (They were also used earlier by Bézout and Euler
[2, p. 46], and at the same time by Vandermonde [22], but Lagrange made
a profound use of them.) Lagrange and others used this method to study
equations of arbitrary degree, see [4, Sections 8.3 and 12.1] and, for exam-
ple, the solutions by Vandermonde and Malfatti of some quintic equations
givin in [2, Chapters 7–8]. This is an important forerunner of Galois the-
ory. In retrospect, the Lagrangre resolvents can perhaps also be seen as the
beginning of discrete Fourier analysis.
We assume, for simplicity, that K ⊆ C, and we then define

1 3
ω := exp(2πi/3) = − + i, (2.13)
2 2
a primitive third root of unity. Note that
ω 3 = 1, 1 + ω + ω 2 = 0, (2.14)
which will be used repeatedly below without further comment.
Remark 2.3. For a general field K, not necessarily contained in C, we can
let ω be a primitive third root of unity in K. It is easily verified that the
formulas below make sense, and are correct, in K, so the result holds in full
generality.
4 SVANTE JANSON

We then define the Fourier transform of a vector (z1 , z2 , z3 ) (regarded as


a function on Z3 ∼
= C3 ) as (bz1 , zb2 , zb3 ), with

zbk := 13 z1 + ω −(k−1) z2 + ω −2(k−1) z3 , (2.15)
and note the Fourier inversion formula, which in this case is easily verified
directly,
zk = zb1 + ω k−1 zb2 + ω 2(k−1) zb3 . (2.16)
Hence, if we define
u := 13 (α1 + ω 2 α2 + ωα3 ) = 13 (β1 + ω 2 β2 + ωβ3 ), (2.17)
v := 13 (α1 + ωα2 + ω 2 α3 ) = 13 (β1 + ωβ2 + ω 2 β3 ), (2.18)
and note that
− 13 b = 13 (α1 + α2 + α3 ), (2.19)
1
0= 3 (β1 + β2 + β3 ), (2.20)
we see that the Fourier transforms of the vectors (α1 , α2 , α3 ) and (β1 , β2 , β3 )
are (− 13 b, u, v) and (0, u, v), respectively. Consequently, the inversion for-
mula (2.16) yields
α1 = − 13 b + u + v, (2.21)
2
α2 = − 13 b + ωu + ω v, (2.22)
2
α3 = − 13 b + ω u + ωv, (2.23)
and, equivalently,
β1 = u + v, (2.24)
2
β2 = ωu + ω v, (2.25)
β3 = ω 2 u + ωv. (2.26)
To solve the equation f (x) = 0, it thus suffices to find u and v.
The objects u and v are elements of the field E(ω), which is the split-
ting field of f (or g) over K(ω). It is thus a Galois extension of K(ω),
and also of the intermediate field K(δ, ω). An element of the Galois group
Gal(E(ω) : K(ω)) maps E into itself (because it fixes K and E is a normal
extension of K), and thus its restriction to E is an element of Gal(E : K).
This defines a group homomorphism Gal(E(ω) : K(ω)) → Gal(E : K), which
is injective because E(ω) is generated by E and K(ω); thus we can regard
Gal(E(ω) : K(ω)) as a subgroup of Gal(E : K). Similarly, Gal(E(ω) : K(δ, ω))
is a subgroup of Gal(E : K(δ)).
Let H := Gal(E(ω) : K(δ, ω)). Then H ⊆ Gal(E : K(δ)) ⊆ A3 , so if H is
not trivial, then H = A3 and H is generated by a cyclic permutation σ with
σ(αk ) = αk+1 (with indices modulo 3). Then, by (2.17)–(2.18), σ(u) = ωu
and σ(v) = ω 2 v. Consequently, σ(u3 ) = u3 , and σ(v 3 ) = v 3 . This implies
that u3 and v 3 are fixed by the Galois group H, and thus u3 and v 3 belong
to the fixed field FixE(ω) (H) = K(δ, ω). We can find them as follows, using
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 5

x3 − 1 = (x − 1)(x − ω)(x − ω 2) and thus x3 − y 3 = (x − y)(x − ωy)(x − ω 2y):


First, by (2.24)–(2.26) and (2.6),
u3 + v 3 = (u + v)(u + ωv)(u + ω 2 v) = β1 β2 β3 = −q. (2.27)
Next, by (2.17)–(2.18),
u − v = 31 (ω 2 − ω)(α2 − α3 ) (2.28)
1
u − ωv = 3 (1 − ω)(α1 − α3 ) (2.29)
2 1 2
u−ω v = 3 (1 − ω )(α1 − α2 ) (2.30)
and thus, using (2.11),

3 3 − 3i
2
u − v = (u − v)(u − ωv)(u − ω v) = (α1 − α2 )(α1 − α3 )(α2 − α3 )
r 9

3i ∆
=− δ= − . (2.31)
9 27
Remark 2.4. The choice of square root in (2.31) is not important, since
a change of sign of it is equivalent to an interchange of u and v, which
just permutes the roots α2 and α3 (β2 and β3 ).) This reflects the fact that
if the Galois group Gal(E(ω) : K(ω)) contains an odd permutation τ , then
τ (u) = ω j v and τ (v) = ω j u for some j = 0, 1, 2; thus τ (u3 ) = v 3 and
τ (v 3 ) = u3 .
We thus find, recalling (2.10),
p r r 
3 −q + −∆/27 q −∆ q p 3  q 2
u = =− + =− + + , (2.32)
2 2 108 2 3 2
p r r 
3 −q − −∆/27 q −∆ q p 3  q 2
v = =− − =− − + . (2.33)
2 2 108 2 3 2
We then find u and v by taking cube roots. In order to choose the right
roots, we also compute, from (2.17)–(2.18) and (2.4)–(2.5),

uv = 91 β12 + β22 + β32 + (ω + ω 2 )(β1 β2 + β1 β3 + β2 β3 )

= 19 (β1 + β2 + β3 )2 − 3(β1 β2 + β1 β3 + β2 β3 ) (2.34)
= − 13 p.
If we replace u by the alternative cube root ωu or ω 2 u, we thus have to
replace v by ω 2 v or ωv, respectively, which by (2.21)–(2.23) and (2.24)–
(2.26) yields a cyclic permutation of the roots α1 , α2 , α3 or β1 , β2 , β3 . We
summarize:
Theorem 2.5 (Cardano’s formula). The roots of g(y) = y 3 + py + q are
given by
s r  s r 
3 q p 3  q 2 3 q p 3  q 2
− + + + − − + , (2.35)
2 3 2 2 3 2
6 SVANTE JANSON

where the two square roots are chosen to be the same, and the two cube roots
are chosen such that their product is −p/3; this gives 3 choices for the cube
roots, which gives the 3 roots of g(y) = 0. (In the exceptional case p = q = 0,
the only root 0 is counted thrice.)
Equivalently, the roots of f (x) = x3 + bx2 + cx + d are given by
s r  s r 
b 3 q p 3  q 2 3 q p 3  q 2
− + − + + + − − +
3 2 3 2 2 3 2
s r s r (2.36)
b 3 q −∆ 3 q −∆
=− + − + + − − ,
3 2 108 2 108
with p and q given by (2.7)–(2.8), and ∆ given by (2.10).
Remark 2.6. This formula is known as Cardano’s formula since it was first
published by Cardano in Ars Magna [3], although it is attributed by him to
Scipione del Ferro, see Appendix B.
Remark 2.7. The case p = q = 0 is exceptional because then (and only
then) u3 = v 3 = 0. This is the trivial case when f and g are cubes (x + b/3)3
and y 3 and thus have triple roots −b/3 and 0, respectively.
The case with a double (but not triple) root are handled correctly by
Theorem 2.5. This is the case when ∆ = 0 (but not p = q = 0), and thus
u3 = v 3 (6= 0). We can find a cube root u = v with uv = u2 = −p/3,
and then the other eligible pairs of cube roots are (ωu, ω 2 u) and (ω 2 u, ωu),
yielding the roots β1 = 2u, β2 = β3 = −u, and thus α1 = −b/3 + 2u,
α2 = α3 = −b/3 − u.
Similarly, there are no problems in the case when u3 or v 3 is 0, but not
both. This happens, by (2.32)–(2.33) and (2.34), when p = 0 but q 6= 0.
Choosing the square root such that v 3 = 0, we have u3 = −q; the polynomial
g(y) equals y 3 + q which has the three roots u, ωu, ω 2 u.
Remark 2.8. By (2.27) and (2.34), which implies u3 v 3 = −p3 /27, u3 and
v 3 are the roots of the quadratic resolvent
r(x) := x2 + qx − p3 /27 ∈ K[x]. (2.37)
Note that the quadratic resolvent has discriminant, by (2.31),
∆ 1
Dis(r) := (u3 − v 3 )2 = − = − Dis(f ). (2.38)
27 27
3. Polynomials of degree 4
Let f (x) = x4 + bx3 + cx2 + dx + e, with b, c, d, e ∈ K, be a polynomial of
degree 4, and let α1 , α2 , α3 , α4 be its roots in some extension of K. Thus
f (x) = x4 + bx3 + cx2 + dx + e = (x − α1 )(x − α2 )(x − α3 )(x − α4 ). (3.1)
It is convenient to make the translation x = y − b/4, converting f (x) into
g(y) := f (y − b/4) = y 4 + py 2 + qy + r (3.2)
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 7

(without third degree term) for some p, q, r ∈ K. Thus g has the roots
β1 , β2 , β3 , β4 with βi = αi + b/4, so αi = βi − b/4, i = 1, . . . , 4. Hence,
g(y) = y 4 + py 2 + qy + r = (y − β1 )(y − β2 )(y − β3 )(y − β4 ). (3.3)
The polynomials f and g have the same discriminant
Y Y
∆ := Dis(f ) = Dis(g) = (αi − αj )2 = (βi − βj )2 . (3.4)
1≤i<j≤4 1≤i<j≤4

Since ∆ is a symmetric polynomial in α1 , . . . , α4 (or β1 , . . . , β4 ), it can be


written as a polynomial in the coefficients of f or g. A well-known calculation
yields, see e.g. [14],
∆ = b2 c2 d2 − 4 b2 c3 e − 4 b3 d3 + 18 b3 cde − 27 b4 e2 − 4 c3 d2
+ 16 c4 e + 18 bcd3 − 80 bc2 de − 6 b2 d2 e + 144 b2 ce2
− 27 d4 + 144 cd2 e − 128 c2 e2 − 192 bde2 + 256 e3 (3.5)
3 2 4 4 2 2 2 3
= −4 p q − 27 q + 16 p r + 144 pq r − 128 p r + 256 r . (3.6)
We also define the square root of ∆:
Y Y √
δ := (αi − αj ) = (βi − βj ) = ∆. (3.7)
1≤i<j≤4 1≤i<j≤4

Again, the sign of δ may change if we permute α1 , . . . , α4 ; the sign is pre-


served by an even permutation but is changed by an odd permutation.
Let E = K(α1 , . . . , α4 ) = K(β1 , . . . , β4 ) be the splitting field of f , or
g, over K, and let G := Gal(E : K) be the Galois group of the extension
E ⊇ K. The elements of the Galois group G permute the roots αi (or βi ),
and G may be regarded as a subgroup of S4 .
S4 has a normal subgroup V consisting of the 4 permutations ι (identity)
and (12)(34), (13)(24), (14)(23). Thus G has a normal subgroup G ∩ V . Let
the fixed field of G ∩ V be F . Then F is a Galois extension of K with Galois
group G/G ∩ V ⊆ S4 /V ∼ = S3 .
Fourier analysis on V is especially simple because every element has order
1 or 2, and thus every character is ±1 (again, see e.g. [7] or [20]). We identify
functions on V by vectors (z1 , z2 , z3 , z4 ), with z1 the value at ι, and define
the Fourier transform of (z1 , z2 , z3 , z4 ) as (b
z1 , zb2 , zb3 , zb4 ) with
zb1 := 21 (z1 + z2 + z3 + z4 ), (3.8)
1
zb2 := 2 (z1 + z2 − z3 − z4 ), (3.9)
1
zb3 := 2 (z1 − z2 + z3 − z4 ), (3.10)
1
zb4 := 2 (z1 − z2 − z3 + z4 ). (3.11)
For V , with our chosen normalization, the Fourier inversion formula takes
the especially simple form bzb = z, i.e., the Fourier transform is its own inverse.
(This is easily verified directly.)
8 SVANTE JANSON

Since β1 +β2 +β3 +β4 = 0, the Fourier coefficient βb1 = 0. For convenience,
we shift the indices and define γi := βbi+1 , i = 0, 1, 2, 3, where thus γ0 =
0. The Fourier transforms of (α1 , α2 , α3 , α4 ) and (β1 , β2 , β3 , β4 ) are thus
(− 12 b, γ1 , γ2 , γ3 ) and (0, γ1 , γ2 , γ3 ), where

γ1 := βb2 := 12 (α1 + α2 − α3 − α4 ) = β1 + β2 = −β3 − β4 , (3.12)


γ2 := βb3 := 12 (α1 − α2 + α3 − α4 ) = β1 + β3 = −β2 − β4 , (3.13)
γ3 := βb4 := 1
2 (α1 − α2 − α3 + α4 ) = β1 + β4 = −β2 − β3 . (3.14)
Permutations in V act on the vectors (regarded as functions on V ) by
translations in V , and thus on the Fourier transforms by multiplying by
characters, which are ±1. In other words, permutations in V act on γ1 , γ2
and γ3 by multiplying by ±1 (as is easily seen directly from (3.12)–(3.14)).
Consequently, if we define
u := γ12 = (β1 + β2 )2 = (β3 + β4 )2 , (3.15)
v := γ22 = (β1 + β3 )2 = (β2 + β4 )2 , (3.16)
w := γ32 2 2
= (β1 + β4 ) = (β2 + β3 ) , (3.17)
then u, v and w are fixed by V ∩ G, and the thus belong to the fixed field
F . We can easily find them explicitly. If σ is any element of the Galois
group G, then σ permutes β1 , . . . , β4 , and it follows from (3.15)–(3.17) that
σ permutes u, v, w. Hence any symmetric polynomial in u, v, w is fixed by
every σ ∈ G, and thus it belongs to K. In particular, this applies to the
coefficients of the polynomial R(x) := (x − u)(x − v)(x − w), which thus has
coefficients in K. Calculations yield the explicit formulas
u + v + w = −2p, (3.18)
2
uv + uw + vw = p − 4r, (3.19)
uvw = q 2 . (3.20)
Hence, u, v, w are the three roots of the cubic resolvent
R(x) := (x − u)(x − v)(x − w) = x3 + 2px2 + (p2 − 4r)x − q 2 ∈ K[x]. (3.21)
Remark 3.1. Note that u−v = (β1 −β4 )(β2 −β3 ), u−w = (β1 −β3 )(β2 −β4 ),
v−w = (β1 −β2 )(β3 −β4 ). Hence the discriminant (u−v)2 (u−w)2 (v−w)2 of
the cubic resolvent R equals the discriminant of g or f given by (3.5)–(3.6).
Having found u, v, w, we take their square roots to find γ1 , γ2 , γ3 . By
(3.20), γ1 γ2 γ3 = ±q. In fact, using (3.12)–(3.14) and β1 + β2 + β3 + β4 = 0,
γ1 γ2 γ3 + q = γ1 γ2 γ3 − (β1 β2 β3 + β1 β2 β4 + β1 β3 β4 + β2 β3 β4 )
= −(β3 + β4 )(β2 + β4 )(β2 + β3 ) + (β2 + β3 + β4 )(β2 β3 + β2 β4 + β3 β4 )
− β2 β3 β4
= 0.
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 9

Hence,
γ1 γ2 γ3 = −q, (3.22)

which provides the information we need on the signs of γ1 , γ2 , γ3 . We then


find β1 , . . . , β4 by taking the (inverse) Fourier transform of (0, γ1 , γ2 , γ3 ). We
summarize the resulting algorithm:

Theorem 3.2. Let g(y) = y 4 + py 2 + qy + r, and define the cubic resolvent


R(x) by (3.21). Let the√roots of R√ by u, v, w√(for example found by Theo-
rem 2.5), and let γ1 := u, γ2 := v, γ3 := w, where we choose the signs
so that γ1 γ2 γ3 = −q. Then the roots of g are given by

β1 = 21 (γ1 + γ2 + γ3 ), (3.23)
β2 = 12 (γ1 − γ2 − γ3 ), (3.24)
1
β3 = 2 (−γ1 + γ2 − γ3 ), (3.25)
1
β4 = 2 (−γ1 − γ2 + γ3 ). (3.26)

The roots of f (x) = x + bx3 + cx2 + dx + e are αi = βi − b/4, i = 1, . . . , 4,


where g(y) := f (y − b/4).

Note that changing the signs of some of γ1 , γ2 , γ3 while still preserving


(3.22) (i.e., changing the sign of exactly two of them), just yields a permu-
tation of β1 , . . . , β4 .

Remark 3.3. The formulas (3.23)–(3.26) were given by Euler [5, §5] in
1733.
√ Euler’s
√ motivation was different. For the cubic, Cardano’s formula is
3
U + V where U := u3 and V := v 3 are roots of the quadratic resolvent
3

(2.37). Further, for a quadratic x2 = a there is the trivial formula a.
Hence,√ Euler √ sought
√ by analogy a formula for the roots of a quartic in the
form A+ B + C, and found a cubic equation for A, B, C by substituting
in g(y) = 0, see [5] for details. (In our notation, A = u/4, B = v/4
and C = v/4, so Euler’s cubic equation is our R(4x) = 0; the difference
from our cubic resolvent equation R(u) = 0 is thus only a trivial matter of
normalization.) √ √ √
Euler [5, §§6–8] proceeded to write the solution as 4 E + 4 F + 4 G (with
E = A2 , F = B 2 , G = C 2 ), and found another cubic equation satisfied by
E, F, G. Euler conjectured that similar formulas existed for higher degrees
too,√and in√ particular
√ that
√ the roots of a fifth degree equation could be found
5 5 5 5
as A + B + C + D, where A, B, C, D were the roots of some fourth
degree resolvent; however, he could not find such a resolvent. Of course,
we know that Euler’s conjecture cannot hold, since 100 years later it was
proved by Abel and Galois that in general there is no solution by radicals
for a fifth degree equation.
10 SVANTE JANSON

Remark 3.4. A simple calculation yields, by (3.23)–(3.26) and (3.18),


1 
β1 β2 + β3 β4 = γ12 − (γ2 + γ3 )2 + γ12 − (γ2 − γ3 )2
4 (3.27)
1  1 
= 2γ12 − 2γ22 − 2γ32 = u − v − w = u + p
4 2
and similarly
β1 β3 + β2 β4 = v + p, (3.28)
β1 β4 + β2 β3 = w + p. (3.29)
Hence, the roots of the cubic resolvent are the three values of βi βj + βk βl − p
for different permutations ijkl of 1234.
For the roots αi of f we have, since αi = βi −b/4 and β1 +β2 +β3 +β4 = 0,
b2 b2
α1 α2 + α3 α4 = β1 β2 + β3 β4 + =u+p+ , (3.30)
8 8
and similarly α1 α3 + α2 α4 = v + p + b2 /8, α1 α4 + α2 α3 = w + p + b2 /8.
Remark 3.5. Another method to solve the quartic equation x4 + bx3 +
cx2 + dx + e = 0, also due to Lagrange [17], is to form (cf. Remark 3.4)
s1 := α1 α2 + α3 α4 , s2 := α1 α3 + α2 α4 , s3 := α1 α4 + α2 α3 , (3.31)
e
and the cubic polynomial R(z) = (z − s1 )(z − s2 )(z − s3 ) with these as roots.
This polynomial can be expressed in the coefficients of the equation as, see
[4, Section 12.1],
e
R(z) = z 3 − cz 2 + (bd − 4e)z − d2 − b2 e + 4ce. (3.32)
By Remark 3.4 and (3.21), R(z)e = R(z − p − b2 /8),
so this is our usual cubic
resolvent in disguise.
e
Having found s1 , s2 , s3 by solving the resolvent equation R(s) = 0, one
notes, for k = 1, 2, 3,
(2γk )2 − 4sk = b2 − 4c, (3.33)
and thus p
γk = ± sk − c + b2 /4, (3.34)
which, recalling (3.22), yields the roots by (3.23)–(3.26).

Appendix A. Real cubic equations


Assume that f (x) = x3 + ax2 + bx + c is a polynomial of degree 3 with
real coefficients.
Then f has always 3 complex roots (not necessarily distinct, and as always
given by Theorem 2.5), but the number of real roots may be smaller. The
following theorem shows that the number of real roots is 1 or 3, and that
the discriminant discriminates between the possible cases.
Theorem A.1. Let ∆ be the discriminant of f given by (2.10); thus ∆ is
real.
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 11

(i) If ∆ > 0, then f has three simple real roots.


(ii) If ∆ < 0, then f has one simple real root, and two non-real complex
roots forming a pair of complex conjugates.
(iii) If ∆ = 0, then f has either one double and one simple real root, or
a triple real root.
Proof. Let, as in Section 2, the 3 roots of f in C be α1 , α2 , α3 . Note that
α1 + α2 + α3 = −b. ∆ is real by (2.10). By (2.9), ∆ = 0 if and only if two
of the three roots coincide, i.e., f has a double or triple root. This root has
to be real, since otherwise its conjugate would be another double or triple
root and f would have at least 4 roots (counted with multiplicity), which is
impossible. If there is a real double root α1 = α2 , then α3 = −b − α1 − α2
is real too. This shows (iii).
Now suppose that ∆ 6= 0; thus f has three distinct simple roots in C.
Since f (x) → −∞ as x → −∞ and f (x) → ∞ as x → ∞, f has at least one
real root by the intermediate value theorem. Further, since the roots are
simple, f changes sign at each root, so f has an odd number of real roots.
Hence f has either 1 or 3 real roots.
If f has 3 distinct real roots α1 , α2 , α3 , then (2.9) yields ∆ > 0.
If f has only one real root, then the roots are α1 , α2 , α2 with α1 ∈ R and
α2 ∈/ R, and (2.9) yields
∆ = (α1 − α2 )2 (α1 − α2 )2 (α2 − α2 )2 = |α1 − α2 |4 (2i Im α2 )2 < 0. 
Remark A.2. For a cubic y 3 +py +q without second degree term, ∆/108 =
−(p/3)3 − (q/2)2 by (2.10), and thus the criterion for case (i) is (p/3)3 +
(q/2)2 < 0; equivalently, p < 0 and |p/3|3 > |q/2|2 . This was found already
by Cardano, see Remark B.4.
The number of real roots is thus easily found. Now let us consider finding
the root(s). There are by Theorem A.1 three cases, which we treat separately
since they turn out to be quite different. (Of course, the roots are always
given by Theorem 2.5, but we now want to perform only real arithmetic, if
possible.)

A.1. ∆ = 0, a double or triple root. In the case ∆ = 0, the roots


are easily found. The double (or triple) root α1 = α2 is also a root of the
quadratic equation f ′ (x) = 0 (choosing the root that also satisfies f (x) = 0),
and then α3 is given by α3 = −b − 2α1 . It is easily p seen that α1 = α2 =
−b/3 + δ, and thus α3 = −b/3 − 2δ, where δ = ± −p/3 with the correct
sign given by sign(δ) = sign(q). (See also Remarks 2.7 and A.5.)

A.2. ∆ < 0, one simple real root. If ∆ < 0, then Cardano’s formula
(2.36) yields the
p unique real root of f by choosing the real cube roots. Note
that −q/2 ± −∆/108 is real and that the product of the real cube roots
in (2.36) is −p/3 as required, because the product is a cube root of (−p/3)3
and −p/3 is real.
12 SVANTE JANSON

A.3. ∆ > 0, three simple real roots: casus irreducibilis. The case
∆ > 0 is much more p complicated. Of course, Cardano’spformula (2.36) still
applies, but now −∆/108 is imaginary and −q/2 ± −∆/108 complex,
so the formula necessarily involves taking cube roots of complex (nonreal)
numbers, even though we know that the final answer p is a real root of f ; the
three different choices of cube roots of −q/2 + −∆/108 lead to the three
different real roots of f (x) = 0.
In this case the imaginary parts thus cancel in (2.36) for any p admissible
choice of cube
p roots. This can also be seen as follows: since −q/2+ −∆/108
and −q/2 − −∆/108 are complex conjugates, we may, and have to, choose
cube roots of them that are complex conjugates in (2.36); recall that the
product of these cube roots has to be −p/3, which is real. Hence, Cardano’s
formula (2.36) for the roots may be written
s r s r s r
b 3 q −∆ 3 q −∆ b 3 q −∆
− + − + + − + = − + 2 Re − + . (A.1)
3 2 108 2 108 3 2 108

Every complex number may be represented by a pair of real numbers,


viz. its real and imaginary parts, but taking the cube root of a complex
number may not be reduced to a combination of real cube roots (or real
square and higher roots) and usual algebraic algebraic operations. In fact,
it can be shown by Galois theory that if f is any polynomial with rational
coefficients (i.e., f (x) ∈ Q[x]) such that f is irreducible over Q and has
positive discriminant, and α is a root of f , then α cannot be expressed
by real radicals; in other words, there does not exist a sequence of field
extensions Q = F0 ⊂ F1 ⊂ F2 · · · ⊂ FN where Fk = Fk−1 [uk ] for some real
uk with unk k ∈ Fk−1 for some positive integer nk , k = 1, . . . , N , and α ∈ FN ;
see [21, Section 8.8] or [4, Section 8.6]. (Here Q may be replaced by any
subfield of R.)
The case ∆ > 0 is known as the casus irreducibilis; in this case, thus the
equation cannot (in general) be solved by radicals using only real numbers
(somewhat paradoxically, since the answers all are real). This case is, by
(2.10), characterized by −4p3 > 27q 2 , or equivalently

p < 0 and 4|p|3 > 27q 2 ≥ 0 (A.2)

or
p < 0 and |p/3|3 > (q/2)2 . (A.3)
An alternative to Cardano’s formula (A.1) in the case ∆ > 0 are the
following trigonometric formulas, which involves only real numbers but use
transcendental functions instead of algebraic expressions.

Theorem A.3. If the real cubic polynomial f (x) = x3 + bx2 + cx + d has


positive discriminant ∆ > 0, or equivalently, (A.3) holds, then f has three
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 13

real roots given by


r   
b −p 1 −q/2
− +2 cos arccos (A.4)
3 3 3 (−p/3)3/2
where different branches of arccos yield the three different roots.
Equivalently, the roots are given by the three different values of
r   
b −p 1 q/2
− +2 sin arcsin . (A.5)
3 3 3 (−p/3)3/2
p p
Proof. Let z := u3 = −q/2 + −∆/108 = −q/2 + i ∆/108, see (2.32).
Then, using (2.10),
 3
2 q2 ∆ p3 −p p 3
|z| = + =− = = .
4 108 27 3 3
We write z in polar coordinates: z = reiϕ , where thus r = |z| = (|p|/3)3/2
and
Re z −q/2 −q/2
cos ϕ = = = .
|z| r (−p/3)3/2
By (A.1), a root of f is given by
b b b
− + 2 Re z 1/3 = − + 2 Re r 1/3 eiϕ/3 = − + 2r 1/3 cos(ϕ/3),
3 3 3
which yields (A.4), with different choices of ϕ yielding the three roots.
To see (A.5), let ψ := ϕ + 3π/2 and note that sin ψ = − cos ϕ and
sin(ψ/3) = cos(ϕ/3). 

In (A.4) and (A.5) we find the different roots by choosing different values
of arccos or arcsin. Often it is more convenient to make a single choice (for
example the principal value with 0 ≤ ϕ ≤ π or −π/2 ≤ ψ ≤ π/2, but any
choice will do).
Theorem A.4. Suppose that the real cubic polynomial f (x) = x3 + bx2 +
cx + d has positive discriminant ∆ > 0, or equivalently, that (A.3) holds.
Then, for any choice of
 
−q/2
ϕ := arccos , (A.6)
(−p/3)3/2
f has three real roots
r r
b −p ϕ b −p  ϕ √ ϕ
− +2 cos , − − cos ± 3 sin . (A.7)
3 3 3 3 3 3 3
Similarly, for any choice of
 
q/2
ψ := arcsin , (A.8)
(−p/3)3/2
14 SVANTE JANSON

f has three real roots


r r  
b −p ψ b −p ψ √ ψ
− +2 sin , − − sin ± 3 cos . (A.9)
3 3 3 3 3 3 3
Proof. The other possible values of arccos in (A.6) are ±ϕ + 2πn, n ∈ Z.
Hence, the three values of the cos in (A.4) are cos(ϕ/3) and

ϕ ± 2π ϕ 2π ϕ 2π 1 ϕ 3 ϕ
cos = cos cos ∓ sin sin = − cos ∓ sin ,
3 3 3 3 3 2 3 2 3
and (A.4) yields (A.7).
Similarly, the other possible values of arcsin in (A.8) are ψ + 2πn and
3π − ψ + 2πn, n ∈ Z, and the three values of the sin in (A.5) are sin(ψ/3)
and

ψ ± 2π ψ 2π ψ 2π 1 ψ 3 ψ
sin = sin cos ± cos sin = − sin ± cos . 
3 3 3 3 3 2 3 2 3
Remark A.5. The formulas (A.4)–(A.9) are meaningful (with real quanti-
ties only), exactly when p < 0 and |q/2| ≤ |p/3|3/2 , i.e., when (A.3) holds
or in the limiting case ∆ = 0 and p 6= 0. The formulas (A.4)–(A.9) are
valid in the latter case p
too, and then yield the roots −b/3 + δ, −b/3 + δ,
−b/3 − 2δ, where δ = ± −p/3 with sign(δ) = sign(q), as found more easily
in Subsection A.1.
Example A.6. Let f (x) = x3 − x, which evidently has the three real roots
0, ±1.
We have b = 0, p = c = −1, q = d =p0, and, by (2.10),
p ∆ = 4 (which is
verified by (2.9)). Hence, u3 = −q/2 + −∆/108 = −1/27 = 3−3/2 i, and
we find the three cube roots
−i
u1 = √ ,
3
√ 
−i  1 3 1 i
u2 = √ − + i = + √ ,
3 2 2 2 2 3
 √ 
−i 1 3 1 i
u3 = √ − − i =− + √ .
3 2 2 2 2 3
Hence (2.35) and (2.36) yield the three roots of f as
u1 + u1 = 0, u2 + u2 = 1, u3 + u3 = −1.
Alternatively,
 we may use the trigonometric formula (A.4). We have
arccos −(q/2) (−p/3)3/2 = arccos 0 = π/2 + nπ, n ∈ Z, and thus the three
roots are
2 π 2 5π 2 9π
√ cos = 1, √ cos = −1, √ cos = 0.
3 6 3 6 3 6
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 15
 
Similarly, (A.5) yields, since arcsin (q/2) (−p/3)3/2 = arcsin 0 = nπ,
n ∈ Z, the three roots as
2 2 2π 2 4π
√ sin 0 = 0, √ sin = 1, √ sin = −1.
3 3 3 3 3
Using (A.7) with ϕ = π/2 we find the same roots given as
2 π 1  π √ π
√ cos , − √ cos ± 3 sin
3 6 3 6 6
while (A.9) with ψ = 0 yields
2 1  √ 
√ sin 0 = 0, − √ sin 0 ± 3 cos 0 = ∓1.
3 3
Example A.7. Let f (x) = x3 − 7x − 6, which has the roots p3, −1, −2.
3
have p = −7, q = −6 and ∆ = 400. Thus, u = −q/2 + −∆/108 =
We p
3 + −100/27 = 3 + 310
3/2 i, and we find the three cube roots

2 1 5 3 1
u1 = −1 + √ i, u2 = − − √ i, u3 = + √ i.
3 2 2 3 2 2 3
Hence, (2.35) and (2.36) yield the three roots of f as
2 Re u1 = −2, 2 Re u2 = −1, 2 Re u3 = 3.
The trigonometric solution (A.4) yields
r r !
7 1 243 2πk
2 cos arccos + , k = 0, 1, 2, (A.10)
3 3 343 3

and it is far from obvious that this yields three integers 3, −2, −1.
Example A.8. Let f (x) = x3 − 7x2 + 14x − 8, which has the roots 1,
2, 4. Then, by
p (2.7)–(2.10), p = −7/3, q = −20/27 and ∆ = 36. Thus,
10
u3 = −q/2 + −∆/108 = 27 + √i3 , and we find the three cube roots

2 1 1 3 5 1
u1 = − + √ i, u2 = − − i, u3 = + √ i.
3 3 6 2 6 2 3
Hence, (2.36) yields the three roots of f as
7 7 7
+ 2 Re u1 = 1, + 2 Re u2 = 2, + 2 Re u3 = 4.
3 3 3
The trigonometric solution (A.4) yields
√  
7 2 7 1 10 2πk
+ cos arccos √ + , k = 0, 1, 2, (A.11)
3 3 3 7 7 3
which, again surprisingly, yields three integers 4, 1, 2.
16 SVANTE JANSON

Example A.9. Cardano saw the problem with the casus irreducibilis and
asked Tartaglia about it, giving x3 = 9x+10 as an example, see Remark B.4.
In this case, thus f (x) = x3 − 9x − 10, sopp = −9, q = −10
√ and ∆/108 =
3 2 3
(−p/3) − (q/2) = 2. Thus, u = −q/2 + −∆/108 = 5 + 2i, and we find
the three cube roots
√ √ √ √ √ √
√ 1+ 6 3− 2 1− 6 3+ 2
u1 = −1 + 2 i, u2 = + i, u3 = − i.
2 2 2 2
Hence, (2.35) and (2.36) yield the three roots of f as
√ √
2 Re u1 = −2, 2 Re u2 = 1 + 6, 2 Re u3 = 1 − 6.

Example A.10. Cardano [3, Chapter XIII] considered also the equation
y 3 = 8y + 3. He saw that y = 3 is one solution (without discussing the
problem of the casus irreducibilis, see Remark B.4). In modern terms he
then found the other two solutions by finding the roots of the quadratic
polynomial (y 3 − 8y − 3)/(y − 3) = y 2 + 3y +
√ 1; he gave a general formula
for this. (The other two solutions are −(3 ± 5)/2; these √are negative, and
Cardano changes the sign and interprets the result (3 ± 5)/2 as the two
positive solutions of x3 + 3 = 8x.)
Let us instead use Cardano’s formula. In this case, p = −8, q = −3 and
∆ = −4p3 − 27q 2 = 1805. Thus,

3
p 3 19 5
u = −q/2 + −∆/108 = + √ i, (A.12)
2 6 3
and we find the three cube roots
√ √ √ √ √
3 5 5−3 9+ 5 5+3 9− 5
u1 = + √ i, u2 = − √ i, u3 = − + √ i.
2 2 3 4 4 3 4 4 3
Hence, (2.35) and (2.36) yield the three solutions of y 3 = 8y + 3 as
√ √
2 Re u1 = 3, 2 Re u2 = −(3 − 5)/2, 2 Re u3 = −(3 + 5)/2.

Example A.11. Bombielli (1550) considered the equation y 3 − 15y − 4 = 0.


Cardano’s formula (2.35) yields the roots as
q q
3 √ 3 √ √ √
2 + −121 + 2 − −121 = 3 2 + 11i + 3 2 − 11i. (A.13)
Bombielli noted that 4 is a root, and showed in a pioneering calculation with
complex numbers that (2 ± i)3 = 2 ± 11i, and thus (A.13) correctly yields
the root (2 + i) + (2 − i) = 4.
The two other cube roots of 2 + 11 i are
√ √ √ √
3 2 3−1 3 −2 3 − 1
u2 = −1 − + i, u3 = −1 + + i.
2 2 2 2

Hence, the three solutions of y 3 − 15y − 4 = 0 are 4 and −2 ± 3.
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 17

A.4. Further comments for real coefficients.


Remark A.12. In the casus irreducibilis we thus always obtain the roots as
complicated expression involving complex cube roots, even when the roots
are, for example, simple integers. (See Examples A.6–A.11 for simple cases.)
Also in the case of a single real root, Cardano’s formula typically yields
complicated expressions (but now involving real roots only) also for simple
integer solutions.
Example A.13. The equation x3 + 6x = 20 [3, Chapter XI] has the real
root x = 2 (and the complex roots −1 ± 3i). Cardano’s formula yields the
root as
q q q q
3 √ 3 √ 3 √ 3 √
10 + 108 + 10 − 108 = 10 + 108 − 108 − 10. (A.14)
√ √
This indeed equals 2 because 108 ± 10 = ( 3 ± 1)3 , but this is far from
obvious and it is hard to see how (A.14) can be simplified without knowing
the answer.
Remark A.14. The trigonometric formulas (A.4)–(A.9) are valid also when
∆ < 0 (and, more generally, for arbitrary complex coefficients with p 6= 0),
but then the angles ϕ and ψ are complex and the formulas less useful.
For real coefficients with ∆ < 0 (the case of a single real root), we can
choose ϕ in (A.6) or ψ in (A.8) purely imaginary (after a change of sign of
the roots and q in the case p < 0 < q), and the formulas (A.7) and (A.9)
can be rewritten with real hyperbolic functions as follows [12].
Theorem A.15. If the real cubic polynomial f (x) = x3 + bx2 + cx + d has
negative discriminant ∆ < 0, then f has one real and two conjugate complex
roots given as follows.
(i) If p < 0 and q/2 < −|p/3|3/2 , then the roots are
r   
b −p 1 −q/2
− +2 cosh arccosh ,
3 3 3 (−p/3)3/2
r       
b −p 1 −q/2 √ 1 −q/2
− − cosh arccosh ± 3 i sinh arccosh .
3 3 3 (−p/3)3/2 3 (−p/3)3/2
(ii) If p < 0 and q/2 > |p/3|3/2 , then the roots are
r   
b −p 1 q/2
− −2 cosh arccosh ,
3 3 3 (−p/3)3/2
r       
b −p 1 q/2 √ 1 q/2
− + cosh arccosh ± 3 i sinh arccosh .
3 3 3 (−p/3)3/2 3 (−p/3)3/2
(iii) If p > 0, then the roots are
r   
b p 1 q/2
− −2 sinh arcsinh ,
3 3 3 (p/3)3/2
18 SVANTE JANSON
r       
b p 1 q/2 √ 1 q/2
− + sinh arcsinh ± 3 i cosh arcsinh .
3 3 3 (p/3)3/2 3 (p/3)3/2
Remark A.16. The three cases in Theorem A.1 can also be seen geomet-
rically by considering the graph of f (or g) as follows (based on [18]).
Denote the stationary points of g, i.e. the roots of g ′ (y) = 0, by ±δ; thus
the stationary points of f are − 3b ± δ. (Note also that f has an inflection
point at (− 3b , q).) Since g ′ (y) = 3y 2 + p, we have
p
δ = −p/3. (A.15)
In particular, δ is either real (p ≤ 0) or imaginary (p > 0); when p < 0, we
choose for convenience the positive square root in (2.11). Let further
 
h := − g(δ) − g(0) = − δ3 + pδ = 2δ3 . (A.16)
(We choose this sign so that h > 0 when δ > 0.) Thus
f (− 3b ± δ) = g(±δ) = g(0) ∓ h = q ∓ h. (A.17)
If δ > 0, then f thus has a local maximum at − 3b − δ with value q + h,
and a local minimum at − 3b + δ with value q − h. Considering the graph of
f , we see that f (x) = 0 then has three real roots if 0 ∈ (q − h, q + h), two
real roots (of which one double) if 0 = q ± h, and one (simple) real root if
0∈ / [q − h, q + h]. We thus see the three different cases in Theorem A.1, with
(i) if h > |q|, i.e. h2 > q 2 , (ii) if h < |q|, i.e. h2 < q 2 , and (iii) if h = |q|, i.e.
h2 = q 2 .
In the limiting case δ = 0 (which entails h = 0), f has no local maximum
or minimum, but a saddle point at − 3b with f (− 3b ) = g(0) = q. In this case
there is a triple root (case (iii)) if q = 0 and otherwise one simple real root
(case (ii)).
If δ is imaginary (and non-zero), then f ′ (x) 6= 0 for all real x, and thus

f (x) > 0 (since this certainly holds for large x); hence f is strictly increasing
and f (x) = 0 has a single, simple root for any q (case (ii)). In this case, h
is imaginary too, so h2 < 0 ≤ q 2 .
We thus see that in all cases, Theorem A.1 holds with case (i) when
h2 − q 2 > 0, (ii) when h2 − q 2 < 0, and (iii) when h2 − q 2 = 0. This is also
confirmed by a simple calculation showing that
∆ = −4p3 − 27q 2 = 108δ6 − 27q 2 = 27(h2 − q 2 ). (A.18)
Using parameters δ and h also simplify the formulas above a little. Since
h2 = 4δ6 = −4p3 /27, (2.32) becomes
p 
u3 = 12 −q + q 2 − h2 (A.19)
so Cardano’s formula (2.36) for the roots of f becomes
q p  q p 
b
− + 12 −q + q 2 − h2 + 12 −q − q 2 − h2 .
3 3
(A.20)
3
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 19

Similarly, in the casus irreducibilis, which now is characterized by h > |q|,


(A.4) and (A.5) can be written
  q    q 
b 1 b 1
− + 2δ cos arccos − = − + 2δ sin arcsin . (A.21)
3 3 h 3 3 h
Remark A.17. Consider the case of a real cubic f (x) = ax3 + bx2 + cx + d
with a single real root (∆ < 0). A geometric way to find the two complex
roots from the graph of f (on R) has been given by e.g. [13; 11]: Let A be
the intersection of the curve y = f (x) and the x-axis (i.e., the real root), and
let ℓ be a tangent from A to the curve. If the tangent point has x-coordinate

x0 , and the slope of the tangent is k, then the complex roots are x0 ± k i.

Appendix B. History1
The solution to cubic equations was first found c. 1515 by Scipione del
Ferro (1465–1526) in Bologna, at least for some cases. At this time, negative
numbers were not used, nor was 0. Thus (in modern notation) x3 + px =
q, x3 = px + q and x3 + q = px, with positive p and q, were regarded
as different types of equations. (The third type was often ignored. We
know today that it always has one negative solution, which generally was
disregarded, and either zero positive solutions or two (casus irreducibilis);
hence Cardano’s formula will never yield a positive solution using real roots
only. The negative solution was treated by Tartaglia and Cardano by, in
modern terms, changing the sign of x, which transforms x3 + q = px to
x3 = px + q.) Scipione del Ferro could solve the first type and possibly the
second (researchers disagree).
Remark B.1. There are 13 types of nontrivial cubic equations with positive
coefficients: x3 +cx = d, x3 = cx+d, x3 +d = cx, x3 = bx2 +d, x3 +bx2 = d,
x3 + d = bx2 , x3 + bx2 + cx = d, x3 + cx = bx2 + d, x3 + bx2 = cx + d,
x3 = bx2 + cx + d, x3 + d = bx2 + cx, x3 + cx + d = bx2 , x3 + bx2 + d = cx.
These are, for example, discussed separately in Cardano’s Ars Magna [3,
Chapters XI–XXIII].
Similarly, Cardano [3, Chapter V] considers three different types of qua-
dratic equations: x2 = bx + c, x2 + bx = c, x2 + c = bx (as did al-Khwarizmi
c. 800, while Brahmagupta in 628 used both zero and negative numbers
and treated all quadratic equations together), and many types of quartic
equations, see Remark B.5 below.
Note that Cardano discusses negative numbers and negative solutions
(called “false solutions”) to equations [3, in particular Chapters I and XXXVII];
1This appendix is largely based on the Foreword (by Oystein Ore) and Preface (by T.
Richard Witmer) to the English translation of Ars Magna [3], the text itself of Ars Magna
[3], Katscher [15; 16], van der Waerden [22], and The MacTutor History of Mathematics
[19] on the Internet (the articles Quadratic, cubic and quartic equations; Scipione del
Ferro; Nicolo Tartaglia; Girolamo Cardano; Lodovico Ferrari; Tartaglia versus Cardan);
much more details can be found in these references. For the history after Cardano’s Ars
Magna [3], see van der Waerden [22].
20 SVANTE JANSON

however, he does not consider negative coefficients (at least not usually, al-
though there are occasional uses in a few examples, for example [3, Chapter
XXXIX, Problem IX]).
Cardano even makes a pioneering tentative use of imaginary numbers
and complex solutions [3, Chapter XXXVII, Rule II], although he clearly
does not understand them and he seems sceptical to his calculation. Com-
plex numbers were introduced in a consistent way somewhat later by Rafael
Bombelli (1526–1572) in his book Algebra (1572), where he also shows how
to work with negative numbers [19, Rafael Bombielli].
Remark B.2. It is claimed in [19, Scipione del Ferro] that the reduction
(2.2) to an equation without quadratic term (which seems trivial to us)
was known at the time of del Ferro, but this seems incorrect, and I rather
believe the claim by [15] that del Ferro considered only such cubics because
the others were too difficult to be solved. See further Remark B.3.
However, del Ferro kept his solution secret. The traditional story is that
he did not tell anyone about it until his deathbed in 1526, when he told the
solution to his student Antonio Maria Fior. (This seems a bit exaggerated,
since his son-in-law Hannibal della Nave much later showed Cardano a note-
book written by del Ferro presenting the solution, but he certainly told very
few.)
Fior let it become known that he could solve cubic equations (without
disclosing the method). This prompted Nicolo Tartaglia (1500–1557) in
Venice to find solutions. He first found a solution to some equations of
the type x3 + bx = d. He claims [16, XIIII p. 12, XXV p. 15, p. 64] that
he found the solution to all such equations in 1530, but he really could
solve (and construct) only special cases, in modern terms having a negative
integer solution. A public contest was held between Fior and Tartaglia in
1535, where each was to solve 30 problems set by the other (within 40 or
50 days); according to himself [16, XXV p. 13], Tartaglia managed to find
the solutions to the two types x3 + px = q and x3 = px + q on 12 and 13
February 1535, only 8 days before the deadline of the contest2, and then
Tartaglia easily won by solving all 30 problems in 2 hours. (Fior’s problems,
which are given in [16, XXXI pp. 29–31], were all of the type x3 + px = q,
which he did not believe that Tartaglia could solve.)
Girolamo Cardano (1501–1576) in Milan then invited Tartaglia, and man-
aged to make him disclose the method (25 March 1539), after Cardano
had promised Tartaglia to keep it secret until Tartaglia had published the
method himself (something Tartaglia never did, preferring to keep it secret
and regretting that he had told Cardano). Cardano worked on the solution
together with his young assistant Lodovico Ferrari (1522–1565), who in 1541
found a solution to quartic equations.

2According to [16, XXXI p. 29], the contest was on 22 February, which yields a dis-
crepancy in the exact dates.
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 21

Cardano found out that the cubic equation had been solved by del Ferro
before Tartaglia, and used this as an excuse to break his promise to Tartaglia
and publish (in 1545) the solutions of cubic and quartic equations in his large
algebra book Ars Magna [3], where they form a major part. (All 13 types
of cubic equations are discussed separately in detail, but only some of the
possible quartic equations, see below.) Cardano introduces the solution of
the cubic equation with:
Scipio Ferro of Bologna well-nigh thirty years ago discovered
this rule and handed it on to Antonio Maria Fior of Venice,
whose contest with Niccolò Tartaglia of Brescia gave Nic-
colò occasion to discover it. He [Tartaglia] gave it to me in
reponse to my entreaties, though withholding the demonstra-
tion. Armed with this assistance, I sought out its demonstra-
tion in [various] forms. This was very difficult. My version
of it follows. [3, Chapter XI]
The publication led to a bitter dispute between Tartaglia and Cardano–
Ferrari. Tartaglia accused in a book [16, XXXIIII p. 42] (1546) Cardano
of breaking an oath to him to keep the solution secret; he also added some
insults against Cardano. This led to a series of equally insulting pamphlets
(6 each) by Ferrari (defending Cardano, who kept a low profile in the dispute)
and Tartaglia (renewing his accusations and insults), and finally to a public
contest between Tartaglia and Ferrari in Milan on 10 August 1548. (Each
posed 62 problems to the other. Ferrari won clearly; Tartaglia left Milan
after the first day of the contest, when he saw that he was losing.)
Remark B.3. del Ferro, Fior and Tartaglia (with the exception x3 +bx2 = d
discussed above) considered only cubics without second degree term, see
Remark B.2, It seems that the reduction (2.2) of general cubic equations
to this case is due to Cardano, who in [3] uses this reduction in 9 of the
10 types with a quadratic term (the exception is x3 + d = bx2 , which is
reduced by the substitution x = d2/3 /y). (Cardano claims in the beginning
of [3] that those things to which he has not attached any name are his
own discoveries. This is of course no proof that this reduction is his own
invention, but it suggests that he regarded the reduction either as his own
contribution or trivial.) Note that Cardano does the reduction separately for
each type and that he does not discuss the reduction in his earlier chapters
on some transformatons of equations. Moreover, he surprisingly does not use
the corresponding reduction for fourth degree equations (see Remark B.5).
Furthermore, Tartaglia did not know this reduction (until he read [3]); note
that Tartaglia himself only mentions cubics without second degree term in
the poem that he later claimed that he gave Cardano with the solution (see
Remark C.1), and that when he claims to have solved x3 + bx2 = d in 1530,
he says that he had not been able to solve x3 + bx2 + cx = d [16, XIIII p. 12]
(and there is no indication that he found a solution later).
22 SVANTE JANSON

Remark B.4. Cardano quickly realized the problem with the casus irre-
ducibilis, see Appendix A, and wrote to Tartaglia about it on 4 August 1539
[16, XXXVIII p. 48], giving the correct condition for it (see Remark A.2)
and giving x3 = 9x + 10 as an example (see Example A.9). Tartaglia was
no longer cooperative, but it seems that neither Cardano nor Tartaglia un-
derstood how to handle this case.
Cardano ignores the complications of the casus irreducibilis in Ars Magna
[3]. In [3, Chapter XIII] he solves y 3 = 8y + 3, and claims that he obtains
y = 3 (which clearly is a solution) by his method, which seems to be at best
an oversimplification. (Cf. Example A.10.)
Remark B.5. Cardano lists [3, Chapter XXXIX] 20 types of quartic equa-
tions that he can solve; these are the 10 nontrivial cases without cubic term
(excluding the ones with only even powers of x, which are quadratic equa-
tions in x2 ) and, symmetrically, the 10 nontrivial cases without linear term
(which are reduced to the former by inversion).
Cardano states that these cases “are the most general as there are 67
others”; I do not understand which these 67 other cases are. Moreover, there
are 15 cases with all possible terms (cubic, quadratic, linear and constant),
and 7 additional without quadratic terms; these are not mentioned as far as
I can see.
Cardano gives several examples where quartic equations are solved by
Ferrari’s method (see Appendix D); these examples illustrate 4 of the 10
types without cubic term and 2 of the 10 types without linear term, and it
is clear that the method applies to all 20 types.
There is also a single example of an equation with both linear and cubic
terms (x4 + 2x3 = x + 1, [3, Problem XXXIX.XIII]), but this is solved by
special argument reducing this equation to a succession of two quadratic
equations (the√equation implies (x(x + 1))2 = x(x + 1) + 1 so x(x + 1) is the
golden ratio ( 5 + 1)/2).
Note that Cardano [3] does not use the general reduction (3.2) to eliminate
the cubic term (in analogy with his treatment of cubic equations), which,
together with Ferrari’s method, would have given the solution of all types
of quartic equations. I do not know whether this reduction, and thus the
solution to general quartics, was found by Cardano, Ferrari or someone else.

Appendix C. del Ferro’s solution of the cubic equation


Of course, del Ferro, Tartaglia and Cardano did not know Galois theory
when they found the solution in Theorem 2.5. Their method is more direct,
and consists in observing (by a stroke of genius) that if y = u + v, then
y 3 = (u + v)3 = u3 + v 3 + 3uv(u + v) = u3 + v 3 + 3uvy; (C.1)
hence, if we can find two numbers u and v such that
u3 + v 3 = −q (C.2)
3uv = −p, (C.3)
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 23

then y 3 = −py − q, so y is a root of g(y) = 0. Note that (C.2)–(C.3) are the


same as (2.27) and (2.34). To find u and v, we multiply (C.2) by u3 and
substitute (C.3), yielding
u6 + qu3 + (−p/3)3 = 0. (C.4)
This is a quadratic equation in u3 , which is readily solved and yields (2.32);
then u is found by taking the cube root and v is found from (C.3). We see
that this yields the same u and v as the argument in Section 2. (In particular,
(2.33) holds, which shows that choosing the other root in (C.4) just means
interchanging u and v, which does not change the root u + v; this should
be no surprise, since u and v have identical roles in the ansatz y = u + v.)
Consequently, this straightforward method yields the same solution u + v
as given in (2.24) and (2.35), and we obtain another proof of Theorem 2.5.
(To see that the three different choices of u as a cube root of u3 really yield
the three different roots of g(y) = 0, with correct multiplicities if there is a
double root, is perhaps less obvious by this method. We do not give a direct
proof since we already know from Section 2 that this indeed is the case.)
Remark C.1. Actually, the method just described, with y = u + v, is
Tartaglia’s and Cardano’s (and possibly del Ferros’s) version for the equation
y 3 = cy + d (with c, d > 0) [3, Chapter XII] which corresponds to our p < 0,
q < 0. For the equation y 3 + cy = d, which corresponds to our p > 0, q < 0,
del Ferro, Tartaglia and Cardano instead set y = u − v [3, Chapter XI],
using
y 3 = (u − v)3 = u3 − v 3 − 3uv(u − v) = u3 − v 3 − 3uvy, (C.5)
and then find u and v such that u3 − v 3 = −q and 3uv = p. This just means
changing the sign of v in the equations above, which of course yields the
same final result. (But it keeps u and v positive in both cases.)
The third case without second degree term, y 3 + d = cy is reduced by
Cardano to the case y 3 = cy +d [3, Chapter XIII], essentially by substituting
−y for y, although Cardano expresses this differently.
According to Tartaglia [16, XXXIIII pp. 42–43], he gave these rules 25
March 1539 to Cardano in form of the following poem (English translation
from [19, Tartaglia versus Cardan]):
When the cube and things together
Are equal to some discreet number,
Find two other numbers differing in this one.
Then you will keep this as a habit
That their product should always be equal
Exactly to the cube of a third of the things.
The remainder then as a general rule
Of their cube roots subtracted
Will be equal to your principal thing
In the second of these acts,
24 SVANTE JANSON

When the cube remains alone,


You will observe these other agreements:
You will at once divide the number into two parts
So that the one times the other produces clearly
The cube of the third of the things exactly.
Then of these two parts, as a habitual rule,
You will take the cube roots added together,
And this sum will be your thought.
The third of these calculations of ours
Is solved with the second if you take good care,
As in their nature they are almost matched.
These things I found, and not with sluggish steps,
In the year one thousand five hundred, four and thirty.3
With foundations strong and sturdy
In the city girdled by the sea.
The Italian original (which rhymes in the form terza rima) is [15]:
Quando chel cubo con le cose appresso
Se agguaglia qualche numero discreto
Trouan dui altri differenti in esso.
Dapoi terrai questo per consueto
Che’llor produtto sempre sia eguale
Alterzo cubo delle cose neto,
El residuo poi suo generale
Delli lor lati cubi ben sottratti
Varra la tua cosa principale.
In el secondo de cotestiatti
Quando che’l cubo restasse lui solo
Tu osseruarai quest’altri contratti,
Del numer farai due tal part’à uolo
Che l’una in l’altra si produca schietto
El terzo cubo delle cose in stolo
Delle qual poi, per communprecetto
Torrai li lati cubi insieme gionti
Et cotal somma sara il tuo concetto.
El terzo poi de questi nostri conti
Se solue col secondo se ben guardi
Che per natura son quasi congionti.
Questi trouai, & non con paßi tardi
Nel mille cinquecentè, quatroe trenta
Con fondamenti ben sald’è gagliardi
3Venice reckoned the year from 1 March, so February 1735 was still 1734 in Venice [15].
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 25

Nella citta dal mar’intorno centa.


Remark C.2. An equivalent, and somewhat quicker, way to obtain Car-
dano’s formula is to use Viète’s substitution y = u−p/(3u) in y 3 +py +q = 0,
which yields (C.4) directly. (This is obviously equivalent to setting y = u+v
with 3uv = −p as above. See [22, Chapter 3] for Viète’s version of this.)

Appendix D. Ferrari’s solution of the quartic equation


Consider again a fourth degree polynomial g(y) = y 4 + py 2 + qy + r as
in (3.3). The solution to the equation g(y) = 0 given in Theorem 3.2 is not
the solution originally found by Ferrari and presented by Cardano in Ars
Magna [3, Chapter XXXIX] (cf. Appendix B).
Ferrari’s method is as follows (in a modern version). From y 4 + py 2 + qy +
r = 0 we obtain, for any z,
(y 2 + z)2 = y 4 + 2y 2 z + z 2 = (2z − p)y 2 − qy + z 2 − r. (D.1)
We let z := (p + u)/2 and obtain, for any u,
 p + u 2 (p + u)2
y2 + = uy 2 − qy + − r. (D.2)
2 4
The right-hand side is a quadratic polynomial in y, and its discriminant is
 
2 (p + u)2
q − 4u − r = −u3 − 2pu2 − p2 u + 4ru + q 2 = −R(u), (D.3)
4
where R is the cubic resolvent (3.21). Hence, if we choose u as a non-
zero root of R, then the right-hand side of (D.2) √
is the square of a linear
polynomial. More precisely, if we further let γ = u, then the right-hand
side of (D.2) is
q2  q 2  q 2
uy 2 − qy + =u y− = γy − , (D.4)
4u 2u 2γ
and thus (D.2) yields
 p + u 2  q 2
y2 + = γy − . (D.5)
2 2γ
Consequently,
p+u  q 
y2 + = ± γy − . (D.6)
2 2γ
This yields a pair of quadratic equations in y, whose solutions are the four
roots of g(y) = 0. (It thus suffices to choose one non-zero root of R(u) = 0 in
order to find all roots of g(y) = 0. See Remark D.2 below for a justification.)
Remark D.1. Ferrari and Cardano considered, as said above, only equa-
tions with positive coefficients (putting some of them on the right-hand side),
so they used different versions of the method for different signs of our p, q
and r, but the versions are essentially the same.
26 SVANTE JANSON

Moreover, in the original version, first y 4 + py 2 is completed to a square


(usually, at least), yielding
 p 2 p2
y2 + = −qy + − r; (D.7)
2 4
then this is further modified by considering (y 2 + p/2 + t)2 and choosing t
so that the right-hand side becomes a square. This is obviously equivalent
to the one-step completion of a square above, with z = p/2 + t. We further
made the substitution t = u/2 in order to obtain the same form of the cubic
resolvent as before.
See [10] for a detailed study of Cardano’s solutions to quartics.
We can connect Ferrari’s method and the methods in Section 3 as follows,
using the notation in Section 3. The roots β1 , β2 are 12 γ1 ± 21 (γ2 + γ3 ), and
are thus the roots of the quadratic equation, using (3.16)–(3.18) and (3.22)
and assuming γ1 6= 0,
(2y − γ1 )2 = (γ2 + γ3 )2 = v + w + 2γ2 γ3 = −2p − u − 2q/γ1 . (D.8)

This equation can be rewritten, since γ1 = u,
4y 2 − 4yγ1 + 2u + 2p + 2q/γ1 = 0, (D.9)
2
4y + 2u + 2p = 4γ1 y − 2q/γ1 , (D.10)
u+p q
y2 + = γ1 y − . (D.11)
2 2γ1
The other two roots β3 , β4 are obtained by replacing γ1 by −γ1 , the other
square root of u. √
We have thus obtained the equations (D.6), with γ = γ1 := u.
Remark D.2. This derivation of (D.6) from Theorem 3.2 shows clearly that
the two roots of each of the two quadratic equations in (D.6) together yield
the four different roots of g(y) = 0. Typically, the four roots are distinct
and we obtain all roots once each, but even when g has multiple roots and
there are repetitions in the roots of (D.6), we obtain the roots of g with
correct multiplicities from (D.6).
Remark D.3. We started above with a reduced quartic y 4 +py 2 +qy +r (as
did Cardano and Ferrari), but, as noted by Lagrange [17, no. 27] the method
can also be applied directly to a general quartic f (x) = x4 +bx3 +cx2 +dx+e
by expanding (x2 + 2b x + t)2 and using f (x) = 0 in analogy with (D.1); we
then continue as above, obtaining a cubic resolvent equation for t, etc., see
[17, no. 27] or [4, Section 12.1.C] for details. The resolvent equation for t
becomes
c bd − 4e (4c − b2 )e − d2
t3 − t2 + t+ = 0. (D.12)
2 4 8
Comparing with (D.1), and recalling y = x + b/2, we have t = z + b2 /16 =
u/2 + p/2 + b2 /16 and thus u = 2t − p − b2 /8, so the resulting cubic resovent
equation (D.12) is R(2t − p − b2 /8) = 0, with R given by (3.21). Using
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 27

e
Remark 3.5, this can be written as R(2t) = 0, as also follows from (D.12)
and (3.32), so the roots of this resolvent equation are simply si /2, i.e.,
1
2 (αi αj + αk αl ) for permutations ijkl of 1234.

Remark D.4. Expressed in the roots βi , we have by Remark 3.4


z = (u + p)/2 = (β1 β2 + β3 β4 )/2. (D.13)
This also follows by (D.11), which implies β1 β2 = (u + p)/2 + q/2γ1 and,
replacing γ1 by −γ1 , β3 β4 = (u + p)/2 − q/2γ1 .
Remark D.5. Ferrari’s method has the following geometric interpretation
in algebraic geometry, see [6] and [1] for details.
Let w := y 2 . Then (y, w) is a simultaneous solution of w2 +pw+qy +r = 0
and w − y 2 = 0, and thus also of the linear combination
w2 + pw + qy + r + u(w − y 2 ) = 0
for any u. As u varies, this equation defines a family (called pencil) of
quadratic curves (also known as conics) in the (y, w)-plane. A calculation
essentially equivalent to the argument above shows that this conic is singular,
and thus a union of two lines, exactly when R(u) = √ 0, and then the two
lines are w + (p + u)/2 = ±(γy − q/2γ) with γ = u (assuming u 6= 0),
corresponding to (D.6). Ferrari’s method thus can be seen as finding one
singular conic in the pencil and decomposing it into a pair of lines; the
solutions then are given by the intersections between these lines and the
conic w = y 2 .
Remark D.6. Descartes gave in 1637 yet another method to solve quartic
equations (see e.g. [10]). Descartes’ s method is based on trying to factor
g(y) = (y 2 + ky + l)(y 2 + my + n) by identifying the coefficients, which yields
the equations
k + m = 0, km + l + n = p, kn + lm = q, ln = r. (D.14)
This yields m = −k and, after some algebra, R(k2 ) = 0, where R is the
cubic resolvent
√ (3.21). Hence we can solve R(u) = 0, choose one root u, let
k := γ := u and m := −k (the other square root of u); solving for l and n
then yields (for u 6= 0)
 p+u q  2 p+u q 
g(y) = y 2 + γy + − y − γy + + . (D.15)
2 2γ 2 2γ
Consequently, we see again that g(y) = 0 is equivalent to (D.6).

References
[1] D. Auckly, Solving the quartic with a pencil. Amer. Math. Monthly 114
(2007), no. 1, 29–39.
[2] J. Bewersdorff, Algebra für Einsteiger. 2nd ed., Friedr. Vieweg & Sohn
Verlag, Wiesbaden, 2004. English transl.: Galois Theory for Beginners.
A Historical Perspective, AMS, Providence, R.I., 2006.
28 SVANTE JANSON

[3] G. Cardano, Artis magnae, sive de regulis algebraicis. Lib. unus. Qui
& totius operis de arithmetica, quod opus perfectum inscripsit, est in
ordine decimus. (Ars Magna.) Johann Petreius, Nuremberg, 1545. Eng-
lish transl.: The Great Art, or The Rules of Algebra. Translated and
edited by T. R. Witmer, MIT Press, Cambridge, Mass., 1968.
[4] David A. Cox, Galois Theory, Wiley-Interscience, Hoboken, NJ, 2004.
[5] Leonhard Euler, De formis radicum aequationum cuiusque ordinis
coniectatio, Commentarii Academiae Scientiarum Petropolitanae 6
(1738), 216–231. (Presented to the St. Petersburg Academy on Novem-
ber 2, 1733.) English transl.: A conjecture on the forms of the roots of
equations, translated by J. Bell. arXiv:0806.1927.
[6] W. M. Faucette, A geometric interpretation of the solution of the gen-
eral quartic polynomial. Amer. Math. Monthly 103 (1996), no. 1, 51–57.
[7] M. Frazier, An Introduction to Wavelets through Linear Algebra.
Springer, New York, 1999.
[8] D. J. H. Garling, Galois Theory. Cambridge Univ. Press, Cambridge,
1986.
[9] P.A. Grillet, Abstract Algebra. 2nd ed., Springer, New York, 2007.
[10] H. Helfgott and M. Helfgott, A modern vision of the work of Cardano
and Ferrari on quartics. Loci (June 2009), MathDL, The Mathematical
Association of America. DOI: 10.4169/loci003312 .
[11] G. Henriquez, The graphical interpretation of the complex roots of cubic
equations. Amer. Math. Monthly 42 (1935), no. 6, 383–384.
[12] G. C. Holmes, The use of hyperbolic cosines in solving cubic polynomi-
als. Math. Gazette 86 (2002), no. 507, 473–477.
[13] F. Irwin and H. N. Wright, Some properties of polynomial curves. Ann.
Math. 19 (1917), no. 2, 152–158.
[14] S. Janson, Resultant and discriminant of polynomials. Note N5, 2007.
http://www.math.uu.se/∼svante/papers/#NOTES
[15] F. Katscher, How Tartaglia solved the cubic equation. Loci: Conver-
gence 3 (2006).
http://mathdl.maa.org/mathDL/46/?pa=content&sa=viewDocument&nodeId=2433
retrieved August 2, 2010.
[16] F. Katscher, Die kubischen Gleichungen bei Nicolo Tartaglia. Die
relevanten Textstellen aus seinen ,,Quesiti et inventioni diverse“
auf deutsch übersetzt und kommentiert. Verlag der Österreichischen
Akademie der Wissenschaften, Vienna, 2001.
[17] J. L. Lagrange, Réflexions sur la résolution algébrique des équations,
Nouveaux Mémoires de l’Académie royale des Sciences et Belles-Lettres
de Berlin, 1770–1771. Reprinted in Œuvres de Lagrange, vol. 3, pp.
205–421, J.-A. Serret ed., Gauthier-Villars, Paris, 1869
[18] R. W. D. Nickalls, A new approach to solving the cubic: Cardan’s
solution revealed. Math. Gazette 77 (1993), no. 480, 354–359.
[19] J. J. O’Connor & E. F. Robertson, The MacTutor History of Mathe-
matics archive. University of St Andrews, Scotland.
ROOTS OF POLYNOMIALS OF DEGREES 3 AND 4 29

http://www-history.mcs.st-andrews.ac.uk/
[20] J.-P. Serre, Représentations linéaires des groupes finis. 2nd ed., Her-
mann, Paris, 1967.
[21] B. L. van der Waerden, Algebra I (German). 7th ed., Springer-Verlag,
Berlin, 1966. English transl: Algebra, Vol. I, Springer-Verlag, New York,
1991.
[22] B. L. van der Waerden, A History of Algebra. Springer-Verlag, Berlin,
1985.
Department of Mathematics, Uppsala University, PO Box 480, SE-751 06
Uppsala, Sweden
E-mail address: svante.janson@math.uu.se
URL: http://www.math.uu.se/∼svante/